Segmentation fault with Python 3.14.1, 3.13.10: insertdict: Assertion `!_PyDict_HasSplitTable(mp)' failed.

[emontnemery]

Bug report

Bug description:

There's a segmentation fault in Python 3.14.1 and 3.13.10 which doesn't happen in the previous point releases.

I'm sorry about the messy reproducer, the issue was noticed in a much larger program and this is what I ended up with after removing all the cruft.

seg_fault_issue_import.py:

import asyncio
import atexit


class OtherClass:
 __slots__ = 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i'

 def __init__(self):
 pass


def register_async_client_cleanup():

 def cleanup_wrapper():
 loop = asyncio.new_event_loop()
 loop.close()

 atexit.register(cleanup_wrapper)

register_async_client_cleanup()

seg_fault_debug.py:

from enum import StrEnum

from seg_fault_issue_import import OtherClass

class MyClass:

 def __init__(self):
 self.attr = OtherClass()


class MyEnum(StrEnum):
 ATTR = "attr"

mydata = MyClass()
setattr(mydata, MyEnum.ATTR, None)

Running seg_fault_debug.py causes a segmentation fault with Python 3.13.10 and 3.14.1:

$ python3 seg_fault_debug.py
Segmentation fault (core dumped)

Original reproducer

seg_fault_issue_import.py:

import asyncio
import atexit
from dataclasses import dataclass


@dataclass(slots=True)
class CPU:
 """CPU."""

 count: int | None = None
 frequency: float | None = None
 load_average: float | None = None
 per_cpu: list[float] | None = None
 power: float | None = None
 stats: float | None = None
 temperature: float | None = None
 times: float | None = None
 times_percent: float | None = None


def register_async_client_cleanup():

 def cleanup_wrapper():
 loop = asyncio.new_event_loop()
 loop.close()

 atexit.register(cleanup_wrapper)

register_async_client_cleanup()

seg_fault_debug.py:

from dataclasses import dataclass, field
from enum import StrEnum

from seg_fault_issue_import import CPU

@dataclass
class MyDataclass:

 attr: None = field(default_factory=CPU)


class MyEnum(StrEnum):
 ATTR = "attr"

mydata = MyDataclass()
setattr(mydata, MyEnum.ATTR, None)

CPython versions tested on:

3.14

Operating systems tested on:

Linux

Linked PRs

• gh-142218: Fix split table dictionary crash #142229
• [3.14] gh-142218: Fix split table dictionary crash (gh-142229) #142244
• [3.13] gh-142218: Fix split table dictionary crash (gh-142229) #142245

[johnslavik]

CC @ZeroIntensity (mentorship): type-crash, 3.14, 3.13, remove type-bug

I'll try to reproduce

[johnslavik]

Triggers an assertion error in debug mode for 3.14, I assume the same for 3.13:

❯ ./python seg_fault_debug.py 
python: Objects/dictobject.c:1866: insertdict: Assertion `!_PyDict_HasSplitTable(mp)' failed.
Aborted (core dumped)

I'll see if we can shrink the reproducer.

[emontnemery]

I simplified the reproducer a little bit by removing the dataclasses, but it's still messy

[johnslavik]

This is related to the slow attribute resolution path -- can't reproduce with fast slot access, which is perhaps why the assertion error regards a dictionary insertion (__dict__).

[johnslavik]

Smaller, without dataclasses or asyncio:

from enum import Enum

class O:
 def __init__(self):
 self.attr = "whatever"

class E(str, Enum):
 a = "attr"

setattr(O(), E.a, "new-whatever")

("whatever" and "new-whatever" could be any other objects, e.g. None and an empty tuple)

[johnslavik]

Reproduced on 3.15, thanks @ZeroIntensity. This must have been a flawed bugfix backported everywhere.

[johnslavik]

from enum import Enum

class O:
 def __init__(self):
 self.attr = "whatever"

class E(str, Enum):
 a = "attr"

O().__dict__.__setitem__(E.a, "something else") # `E.a` as the string seems to cause segfault